Learning latent conformational landscapes encoded in cryo-EM

Abstract

A protein populates a landscape of structural states, abundant snapshots of which are sampled in every cryo-EM experiment. Current analysis averages these snapshots into single density maps or partitions them into discrete classes, discarding the continuous dynamics encoded in the data. Continuous latent-space methods offer a promising alternative, yet whether their learned representations are physically grounded remains unresolved. Here, we realize cryo-EM as structural landscape microscopy, in which latent density directly reflects the probabilistic distribution of molecular states. A central question is whether such a landscape reflects physical reality. For integrin v{beta}8, the learned landscape shows strong agreement with independently derived molecular dynamics simulations, supporting its physical plausibility. We then apply the landscape to structural states that conventional cryo-EM cannot resolve. For LIS1-mediated dynein activation, the landscape reveals a spectrum of states from dominant conformations to low-population intermediates defined by distinct binding modes, including a previously unresolved state. For the KCTD5/CUL3NTD/G{beta}{gamma} complex, the landscape resolves continuous conformational pathways directly from experimental data. Probability-guided particle selection further improves reconstruction quality.